Detection/OCR: annotation-based training with each family's standard loss, and metric-improvement tests

[ooples]

Summary

The detection and OCR models train only through Train(Tensor<T> input, Tensor<T> expectedOutput): mean-squared error between the raw network output and a tensor of the same shape. That proves the networks are differentiable end to end (#2152), but it is not how any of these models is trained, and it cannot drive the task metrics. Nothing today can show that training improves mAP, H-mean or CER.

This is the last acceptance item of #2152, split out by decision on #2154.

What is missing

A training entry point that takes annotations, not raw-output tensors, and applies each family's standard loss with its standard target assignment:

family	annotation	standard loss (reference)
YOLOv8 / v9 / v11	boxes + classes	task-aligned assignment, BCE cls + CIoU + DFL (YOLOv8, TOOD)
YOLOv10	boxes + classes	consistent dual assignment: one-to-many + one-to-one heads (YOLOv10)
DETR / DINO / RT-DETR	boxes + classes	Hungarian set loss: CE + L1 + GIoU (DETRSetLoss exists)
Faster / Cascade R-CNN	boxes + classes	RPN: anchor IoU sampling, BCE + smooth-L1; RoI head: CE + smooth-L1, per cascade stage with rising IoU thresholds
DBNet	polygons	BCE on shrunk probability map + L1 threshold map + dice on binary map (Liao et al. 2020)
CRAFT	character/affinity heatmaps	MSE on region and affinity Gaussian maps
EAST	quads	balanced BCE score + IoU/angle geometry (Zhou et al. 2017)
CRNN	text	CTC (CTCLoss exists)
TrOCR	text	cross-entropy with teacher forcing

Acceptance

• A public training API on the three family bases taking image(s) plus annotations (boxes/classes, polygons, or text), exposed through the facade.
• Each model uses its paper's loss and target assignment (table above).
• Family fixtures gain an overfit test: training on a small synthetic set (for example one to four images with known boxes, polygons or strings) raises mAP@0.5 / H-mean, or lowers CER, measured with ObjectDetectionMetrics, TextDetectionMetrics and TextRecognitionMetrics, against the untrained model.

🤖 Generated with Claude Code

https://claude.ai/code/session_016jgqTmscEnkgmAp1TkNFpG